Chiz/omniASR-igbo-blindspots

omniASR Igbo Blind Spot Dataset

Research Questions

This dataset investigates three interrelated questions about multilingual ASR performance on tonal languages:

1. Operational Definition: What does "language support" mean when a model lists 1,600+ languages? Does coverage imply functional accuracy on linguistically meaningful distinctions?

2. Diagnostic Validity: Can tonal diacritic preservation serve as a diagnostic for acoustic competence vs. orthographic pattern matching in low-resource languages?

3. Systematic Evaluation: Does facebook/omniASR-CTC-1B exhibit systematic tonal collapse in Igbo, and if so, what error patterns emerge?

Overview

This dataset provides a small, controlled diagnostic evaluation of tone and diacritic handling in facebook/omniASR-CTC-1B when processing Igbo (ibo_Latn), a tonal Niger-Congo language with ~45 million speakers. It has 21 audio clips from one speaker, designed to probe specific failure modes.

Key findings (all rates are % of expected marks; dropped and added marks are counted per clip, then summed):

• On the six tonal clips, 37 of 49 expected marks are missing (drop rate 75.5%, clip-level bootstrap 95% CI [57.1%, 89.7%]). Tone-bearing letters fall from 43 in the references to 5 in the model output.
• The model also adds marks (25 across the dataset, added rate 25.8%). Overall net loss is 26.8%, but overall drop rate is 52.6% and DER is 78.4%, so net loss understates how often the output disagrees with the reference.
• On the akwa and oke minimal-pair clips the output contains no tone marks at all.

These results are consistent with the model relying on text-like priors rather than tone acoustics, but one speaker, 21 clips, a single annotator and one model cannot establish that mechanism. This is a blind-spot probe, not a prevalence estimate.

Code and analysis: https://github.com/chizkidd/igbo-asr-tonal-evaluation

Motivation

Recent work on ASR fairness has documented systematic performance disparities across demographic groups (Koenecke et al., 2020) and low-resource African languages (Nakatumba-Nabende et al., 2025). However, existing evaluations focus primarily on word error rates in high-resource languages. This dataset addresses three critical gaps:

1. Tonal language evaluation: Most ASR benchmarks ignore whether models preserve linguistically meaningful tone distinctions
2. Low-resource African languages: Igbo remains underrepresented in ML evaluation despite being a major world language
3. Native speaker ground truth: References were written and verified by a native Igbo speaker (the recording speaker), which captures tone placement that automated metrics cannot. This is a single annotator, so no inter-annotator agreement is available

The Paradox of "Supported" Languages

omniASR's model card lists Igbo (ibo_Latn) among its 1,600+ supported languages. However, as recent work on low-resource ASR demonstrates, nominal support does not guarantee functional accuracy (see Nigatu et al., 2024, "The Zeno's paradox of 'low-resource' languages").

The challenge is definitional: what does it mean for a language to be "low-resource"?

• By training data: Igbo has fewer hours than English (low-resource)
• By speaker population: 45 million speakers (NOT low-resource)
• By model performance: On this small probe it behaves like a low-resource language for tone marks despite being "supported"

This dataset reveals the gap between coverage (language is in the training set) and competence (model preserves linguistically meaningful distinctions). Nigatu et al. (2024) discuss the risk of creating a Zeno's paradox: models claim to support more and more languages, yet the quality asymptote never reaches parity with high-resource languages.

Our contribution: A small native-speaker-annotated probe that makes this gap measurable for one speaker of Igbo, as a starting point for larger studies.

Dataset Structure

huggingface_dataset/
├── audio/               # 21 WAV files (16kHz mono)
├── metadata.csv         # Ground truth, model outputs, error metrics
└── README.md            # This file

Metadata Schema

Column	Description
file_name	Path to audio file
ground_truth	Correct transcription with tone marks
model_output	omniASR-CTC-1B prediction
category	Error category (see taxonomy below)
subcategory	Specific test condition
language	Language code (ibo_Latn, yor_Latn, fra_Latn, mixed)
character_error_rate	Legacy score, 1 - difflib.SequenceMatcher ratio on lowercased text (0-1). Not Levenshtein CER; see Metric definitions for corrected CER
diacritics_expected	E: number of marked letters in the ground truth (a letter carrying a grave, acute, macron or dot-below; a letter with dot-below plus tone counts once)
diacritics_produced	P: number of marked letters in the model output (same rule)
diacritic_loss	Net difference E - P for the clip (negative = more marks produced than expected). Dropped and added marks are max(0, E - P) and max(0, P - E) per clip

Error Taxonomy

1. Cross-lingual Orthographic Interference (5 samples)

Hypothesis: Model applies incorrect orthographic conventions from other languages to Igbo text.

Tests:

• Personal names (01_script_names)
• Formal greetings (02_script_formal)
• Numeric sequences (03_script_numbers)
• Proverbs (04_script_proverb)
• Prosody variation (05_script_slow)

Finding: Expected 18 marks, produced 25: 2 dropped and 9 added (net loss -38.9%, drop rate 11.1%, added rate 50.0%). Most references here carry only dot-below marks, and 4 of the produced marks are tone marks, so part of the "addition" is orthographic convention rather than tone. The pattern is consistent with interference from other supported languages, but does not demonstrate it.

2. Phonemic Tone Sensitivity (6 samples)

Hypothesis: Model cannot distinguish phonemically contrastive tones in Igbo.

Tests:

• Minimal pairs: akwa/akwà/àkwà/ákwá (06_tonal_akwa)
• Minimal pairs: oke/òkè/ọkè (07_tonal_oke)
• Dense tone marks (08_tonal_dense)
• Monotone control (09_tonal_flat)
• Yoruba controls (10_tonal_yoruba, 21_tonal_yoruba_formal)

Finding:

• Drop rate 75.5% (37 of 49 expected marks dropped), added rate 14.3% (7 added), net loss 61.2%, DER 89.8%
• Clip-level bootstrap 95% CI on the drop rate: [57.1%, 89.7%]
• On the monotone clip (09_tonal_flat) the reference has no marks and the output has 7 (3 tone marks, 4 dot-below letters); its CER of 61.9% is the highest of the 21 clips
• Model outputs collapse multiple tonal minimal-pair forms into a shared orthographic representation, indicating weak tonal separability in this evaluation setup

Linguistic Impact: In Igbo, tone changes word meaning. Losing tone marks is equivalent to losing consonants in English (e.g., "bat" vs "hat" vs "cat" all transcribed as "at").

3. Language Boundary Effects (5 samples)

Hypothesis: English-Igbo code-switching (extremely common in Nigerian speech) disrupts language-specific processing.

Tests:

• English → Igbo embedding (11_codeswitch_en2ig)
• Igbo → English embedding (12_codeswitch_ig2en)
• Sentence-level alternation (13_codeswitch_alternate)
• Diacritics in English context (14_codeswitch_embedded)
• Nigerian Pidgin control (15_codeswitch_pidgin)

Finding: Net loss 14.3%, but 7 of 14 expected marks were dropped (drop rate 50.0%) and 5 were added (added rate 35.7%); the net figure hides both. English portions are mostly transcribed while adjacent Igbo loses marks (e.g., "The ụlọ is beautiful" → "te ulọ is beautiful"), consistent with language-boundary effects on orthographic fidelity, though 5 clips cannot confirm it.

4. Domain-Specific Lexical Coverage (5 samples)

Hypothesis: Model struggles with culturally specific terms, place names, and idiomatic expressions outside training distribution.

Tests:

• Nigerian place names (16_context_places)
• Igbo food terms (17_context_food)
• Long proverbs (18_context_proverb)
• French control (19_context_french)
• Background noise robustness (20_context_noise)

Finding:

• Lowest net loss (6.3%), but 5 of 16 expected marks were dropped (drop rate 31.2%) and 4 added (added rate 25.0%), with high word-level errors (mean CER 37.4%)
• Place names corrupted: "Owerri" → "weri" (missing syllable)
• High-resource French performed unexpectedly poorly (Czech/Slavic character hallucinations)

Sample Audio Examples

You can listen to individual audio files in the Files tab or explore the full dataset with metadata in the Dataset Viewer.

Key examples:

• 06_tonal_akwa.wav - Tonal minimal pairs (4 different words collapsed to random outputs)
• 09_tonal_flat.wav - Monotone speech with marks added
• 11_codeswitch_en2ig.wav - Code-switching (English perfect, Igbo loses tones)

Example 1: Tonal Minimal Pairs (06_tonal_akwa.wav)

• Ground Truth: akwa, akwa, akwa. Akwà, akwà, akwà. Àkwà, àkwà, àkwà. Ákwá, ákwá, ákwá.
• Model Output: akua akua akua akua akwa akwa akwa akua akwa ọkua ọkua ọkua
• Error: Model collapses 4 distinct words into random variations

▶Listen to audio

Example 2: Monotone Hallucination (09_tonal_flat.wav)

• Ground Truth: O na-eri oji n'ututu. (spoken with FLAT intonation, no tones)
• Model Output: ọne rị ọjí nụ tútú (model ADDED diacritics that weren't spoken)
• Observation: The output carries 3 tone marks and 4 dot-below letters that the reference does not have. Consistent with text priors, but this is one clip, and the dot-below additions may simply be standard orthography (the reference was written without dots)

▶Listen to audio

Quantitative Summary

E = expected marks, D = dropped, H = added (both per clip, then summed).

Category	Clips	E	D	H	Net loss	Drop rate	Added rate	DER	Mean CER
Phonemic Tone Sensitivity	6	49	37	7	61.2%	75.5%	14.3%	89.8%	51.6%
Cross-lingual Orthographic Interference	5	18	2	9	-38.9%	11.1%	50.0%	61.1%	28.5%
Language Boundary Effects	5	14	7	5	14.3%	50.0%	35.7%	85.7%	25.7%
Domain-Specific Lexical Coverage	5	16	5	4	6.3%	31.2%	25.0%	56.2%	37.4%
Overall	21	97	51	25	26.8%	52.6%	25.8%	78.4%	36.5%

An earlier version of this card listed net loss for three categories and drop rate for the tonal row in one "Diacritic Loss" column. Both are shown separately above.

Statistical Analysis

Notation and metric definitions

For each clip i: E_i expected marks, P_i produced marks, D_i = max(0, E_i - P_i) dropped, H_i = max(0, P_i - E_i) added. Totals sum over clips, so additions in one clip never cancel drops in another.

• Net loss = (E - P) / E (can be negative)
• Drop rate (RDD) = D / E
• Added rate = H / E
• DER = (D + H) / E (can exceed 100%)
• Share of produced marks that are additions = H / P (earlier called "hallucination rate")
• CER = Levenshtein distance / reference length, lowercased NFC text. The character_error_rate column is a different, similarity-based score (see schema).

A "mark" is one letter (NFD cluster) carrying a grave, acute, macron or dot-below. This rule reproduces diacritics_expected and diacritics_produced for all 21 rows.

Results with uncertainty

Bootstrap: 10,000 resamples of clips (the independent unit; resampling individual marks would give intervals that are too narrow), percentile 95% CI. The headline is the point estimate on the original data, not the bootstrap mean.

Metric	Group	Point	95% CI
Drop rate	Phonemic Tone Sensitivity	75.5%	[57.1%, 89.7%]
Drop rate	Overall	52.6%	[30.3%, 69.7%]
Drop rate	Language Boundary	50.0%	[0%, 91.7%]
Drop rate	Domain-Specific	31.2%	[0%, 83.3%]
Drop rate	Cross-lingual	11.1%	[0%, 27.3%]
Share of produced marks that are additions (H/P)	Overall	35.2%	[18.2%, 53.3%]
Share of produced marks that are additions (H/P)	Cross-lingual	36.0%	[8.7%, 68.0%]
Mean CER	Overall	36.5%	[31.1%, 42.1%]
Mean CER	Phonemic Tone Sensitivity	51.6%	[48.4%, 56.0%]

With 5 or 6 clips per category the non-tonal intervals are very wide. Earlier versions reported the legacy similarity score as CER (overall 0.333 [0.267, 0.402], tonal 0.506 [0.416, 0.617]).

Comparing categories

Overlap of two separate intervals is not a test, so categories are compared with a paired bootstrap of the difference in drop rate:

Tonal minus	Difference	95% CI
Cross-lingual	64.4 points	[40.5, 85.7]
Domain-Specific	44.3 points	[-9.4, 80.6]
Language Boundary	25.5 points	[-18.9, 78.6]

Only the contrast with the cross-lingual category excludes zero. The data support "tonal clips drop many marks" but not "tonal clips drop more than every other category".

Position-aware estimate

Counts cannot tell a correct mark from a mark on the wrong letter. Aligning reference and output on base letters and classifying each mark gives, overall, 33 correct, 7 wrong, 57 missing and 31 inserted (tonal clips: 4 correct, 7 wrong, 38 missing, 8 inserted). This is an estimate: a mis-heard letter can make a mark look missing.

Tone-only counts

Dot-below letters (ọ, ị, ụ) are partly an orthography choice, and some references omit them. Counting only grave, acute and macron: the references have 45 tone-bearing letters and the model output 9 (tonal clips: 43 vs 5; cross-lingual clips: 0 vs 4).

Interpretation

The drop rate in the tonal category is high even at the lower bound of its interval, so the loss is unlikely to be an artifact of which six clips were drawn. The interval describes noise in this dataset only: six clips from one speaker are not a random sample of Igbo speech.

Visualizations

Marks dropped vs. marks added, by category

Figure 1: Drop rate (dark) and added rate (grey) per category, as a percentage of expected marks, with counts. Net loss would have hidden the large additions in the cross-lingual and language-boundary categories.

CER vs. drop rate

Figure 2: Per-clip CER against drop rate for the 18 clips with at least one expected mark (Pearson r = 0.36). Tonal clips (red) combine high CER (about 48% to 53%) with 50% to 100% mark loss. Rates sit at 0% or 100% because each clip carries only a few marks. With so few clips, tone-mark loss cannot be cleanly separated from general transcription errors.

Bootstrap confidence intervals

Figure 3: Drop rate with 95% bootstrap intervals (10,000 resamples of clips). Intervals stop at 0% because a rate cannot be negative. The non-tonal intervals are wide because each category has only 5 clips.

Scope and Limitations of Claims

This study demonstrates:

• High mark loss in omniASR-CTC-1B on tonal Igbo clips (21 controlled samples, one speaker)
• Failure to preserve tonal minimal pair distinctions in this evaluation setup
• Marks added on a monotone clip, consistent with text-like priors (one clip; dot-below additions may be standard orthography)

This study does NOT claim:

• That tonal clips lose more marks than every other category (only the contrast with the cross-lingual category is clearly separated)
• That the annotations are validated by more than one person (single annotator)
• That omniASR fails universally on all Igbo speech
• That tone modeling is architecturally absent from the model
• That Igbo is uniquely disadvantaged relative to all other low-resource languages
• That the observed error rates generalize to all dialects or all speakers

What would be needed to strengthen these claims:

• Multi-speaker evaluation (N=10+ speakers across dialects)
• Acoustic analysis (F0 contour extraction, pitch tracking validation)
• Comparative evaluation on other tonal African languages
• Controlled resynthesis experiments isolating acoustic vs. lexical priors

A Lead on Weak Tonal Conditioning: The Monotone Clip

The most suggestive single clip is File 09 (monotone speech):

• Setup: I spoke "O na-eri oji n'ututu" with deliberately flat intonation (no tonal variation); the reference has no marks
• Expected: If marks were tightly conditioned on acoustics, the output would contain few or no marks
• Result: "ọne rị ọjí nụ tútú" has 7 marks: 3 acute tone marks (jí, tú, tú) and 4 dot-below letters

Interpretation: Consistent with mark insertion driven by lexical or orthographic priors, but it is one clip, and the dot-below letters may be correct orthography since the reference was written without dots. Confirming the mechanism would require acoustic analysis (e.g., F0 contour statistics) and controlled resynthesis experiments.

Linguistic Error Analysis: When Tone Loss Changes Meaning

File	Ground Truth	Model Output	Observation
06_tonal_akwa	4 tone patterns on akwa	akua / akwa / ọkua	The four tonal forms are not distinguished; output carries no tone marks
07_tonal_oke	3 tone patterns on oke	oke / oki	The three tonal forms are not distinguished; output carries no tone marks
08_tonal_dense	ọ̀jị̀ (kolanut)	ọjị	Tone marks lost, base word kept
16_context_places	Owerri (city)	weri	Unrecognizable as place name

Meaning glosses for the minimal-pair forms were removed from this table pending verification against a standard Igbo reference.

Impact: Dropping tone can turn one word into another valid word, or into an ambiguous form, which matters for downstream uses such as voice assistants and translation.

Performance Gap: Claimed vs. Measured

According to Meta's omnilingual ASR paper (arXiv:2511.09690), omniASR reports CER below 10% for 78% of supported languages, and Igbo (ibo_Latn) is among the 1,600+ supported languages.

This dataset (true CER, lowercased):

• Overall mean CER: 36.5% (95% CI [31.1%, 42.1%])
• Tonal category: 51.6%
• Highest single clip: 61.9% (09_tonal_flat)

These numbers are not directly comparable to the published ones. This is a 21-clip diagnostic set deliberately built to stress tone, code-switching, names and noise, recorded by one speaker, and scored with a stricter, tone-sensitive reference. The published figures come from different test sets and protocols. The gap is consistent with either (a) Igbo sitting low among supported languages, or (b) benchmark test sets that do not capture tonal accuracy, but this dataset cannot separate those explanations from data-domain or protocol differences.

Implications for Low-Resource ASR

This dataset reveals that raw multilingual coverage (1,600+ languages) does not guarantee linguistic accuracy:

1. Tonal languages require specialized evaluation: WER/CER metrics miss semantic errors when tones are lost. In this dataset, tonal clips lose most of their marks even though the language is listed as supported.

2. Native speaker validation is essential: Automated metrics can miss cases where one tonal form of a word is transcribed as another. We provide single-speaker ground truth as a first baseline before any multi-speaker evaluation (single annotator, so tone placement is unvalidated by others).

3. Code-switching is not a solved problem: Real-world multilingual speech patterns break current ASR systems. Nigerian English-Igbo code-switching represents a common speech pattern that production systems must handle.

4. "Supported" ≠ "Works well": Nigatu et al. (2024) discuss how models can list languages in their documentation while providing functionally inadequate service. Our results indicate a substantial gap between nominal language coverage and functional performance on tone-sensitive orthography in Igbo.

Broader Implications: Linguistic Fairness in Multilingual ASR

Tonal diacritics in Igbo encode phonemic distinctions that alter lexical meaning. Systematic loss of these distinctions has implications beyond transcription accuracy.

Epistemic Distortion

When tonal contrasts are consistently omitted, model outputs may:

• Collapse minimal pairs
• Introduce semantic ambiguity
• Normalize orthographic forms that diverge from standard tonal marking

Such distortions risk misrepresenting core structural features of the language.

Downstream System Impact

ASR increasingly serves as infrastructure for voice interfaces, accessibility tools, translation systems, and educational applications. Tone collapse can propagate semantic errors into downstream systems, particularly in contexts where tonal contrasts determine lexical identity.

Framing

This dataset does not claim quantified downstream harm. Rather, it provides a controlled diagnostic demonstrating that tonal fidelity can degrade substantially under current multilingual ASR evaluation protocols. Future work is required to measure real-world behavioral or allocative consequences.

Comparison to Related Work

Study	Focus	Key Finding
Koenecke et al. (2020)	Racial disparities in commercial ASR	2x higher WER for Black speakers
Nakatumba-Nabende et al. (2025)	Systematic review of bias evaluation and mitigation in ASR for low-resource African languages	Survey of the literature, not a measurement on Igbo tone
Liang & Levow (2025)	Fine-tuning ASR for extremely low-resource fieldwork languages	MMS does best with very little data; XLS-R catches up beyond about 1 hour
This work	Tone and diacritic handling in Igbo ASR	75.5% drop rate on marks in tonal clips (6 clips, one speaker)

Use Cases

This dataset is designed for:

• ASR developers: Benchmark tonal accuracy for African languages
• Linguists: Document systematic biases in multilingual models
• ML fairness researchers: Extend demographic fairness analysis to linguistic fairness
• African NLP community: Provide native-speaker ground truth for Igbo

Recording Methodology

• Speaker: Native Igbo speaker (Nigerian)
• Dialect: Afikpo Igbo (Ebonyi State). Speaker grew up in multilingual Northern Nigerian environment; both parents from Afikpo. Recordings reflect a single-speaker variety and are not intended to represent all Igbo dialects.
• Device: iPhone SE 2nd Generation Voice Memos app
• Format: M4A (AAC codec) converted to 16kHz mono WAV
• Duration: 4-15 seconds per sample
• Environment: Quiet indoor setting (File 20 includes controlled background noise)
• Speech style: Natural conversational pace unless otherwise noted (File 05 is deliberately slow)

Single-speaker recordings are a first baseline before any multi-speaker, multi-dialect evaluation.

Model Details

• Model: facebook/omniASR-CTC-1B
• Features: ASR (Automatic Speech Recognition)
• Parameters: 975,065,300 (~975M)
• Download Size: 3.7 GiB (FP32)
• Inference VRAM: ~3 GiB
• Architecture: CTC-based ASR (wav2vec2-style encoder with CTC head)
• Training: Multilingual (1,600+ languages) on clean, spontaneous speech
• Release: November 14, 2025
• License: Apache 2.0

Reproducibility

All transcriptions generated using:

from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline
pipeline = ASRInferencePipeline(model_card="omniASR_CTC_1B")
transcription = pipeline.transcribe(inp=[audio_path], lang=["ibo_Latn"])

Recomputing the tables: counts, per-clip dropped/added marks, CER, bootstrap intervals, paired contrasts and aligned-mark estimates are all computed from metadata.csv by src/evaluate.py and analysis-2.ipynb in the code repository (CPU only; pandas, numpy, matplotlib, seaborn). The inference run is in analysis.ipynb.

Environment (inference run):

• Google Colab (NVIDIA Tesla T4, 15GB VRAM)
• omnilingual-asr==0.1.0
• torch==2.1.0
• Python 3.12
• Date: March 1, 2026

Limitations and Scope

This dataset represents a proof-of-concept demonstration of native-speaker auditing for low-resource ASR. By design, it prioritizes:

1. Depth over breadth: 21 carefully designed samples targeting specific failure modes rather than 1000s of random utterances
2. Native-speaker authority: Single speaker provides unambiguous ground truth for initial blind spot discovery
3. Systematic coverage: Four distinct categories of errors (orthographic, tonal, code-switching, lexical)

Known limitations:

• Single annotator: All references, including tone placement, were written by one person; no inter-annotator agreement was measured, and some references omit dot-below marks that standard orthography would include
• Clip independence: Clips from one speaker, device and session are not fully independent draws, so even clip-level intervals are optimistic
• Generalizability: Single speaker limits claims about model performance across all Igbo speakers
• Dialectal coverage: Afikpo Igbo only; other varieties (Owerri, Onitsha, Enugu, Nsukka, etc.) are not tested
• Real-world conditions: Primarily clean audio; limited noise robustness testing
• Sample size: 21 recordings establish blind spot existence but not prevalence rates

Why this scope is appropriate: In the spirit of earlier ASR bias audits (Koenecke et al., 2020; Harris et al., 2024), a small controlled set is a reasonable first step for discovering failure modes before any large-scale evaluation. This dataset serves as the foundation for future multi-speaker, multi-dialect studies.

Future Work: Research Agenda

Phase 1: Scale Current Approach (3-6 months)

• Add a second annotator for the reference transcripts (agreement on tone placement)
• Record 50+ samples per category (total: 200+ recordings)
• Recruit 10 speakers across major dialects (Owerri, Onitsha, Enugu, Nsukka, Afikpo)
• Add female/male speaker balance
• Test age range effects (youth vs. elders)

Phase 2: Comparative Model Evaluation (6-12 months)

Audit the same test set on:

• OpenAI Whisper (large-v3)
• Meta MMS (1B-all)
• Google USM
• Microsoft Azure Speech

Research question: Is the high tonal drop rate (75.5% here) specific to omniASR or common across multilingual ASR?

Phase 3: Intervention Studies (12-18 months)

Fine-tuning for low-resource languages is studied by, for example, Liang & Levow (2025):

• Fine-tune omniASR on Igbo data with tonal annotations
• Measure pre/post diacritic accuracy
• Publish open-source fine-tuning pipeline for other tonal African languages

Phase 4: Downstream Impact (18-24 months)

• Partner with Nigerian voice assistant developers
• Measure real-world consequences of tonal errors in deployed systems
• User studies: Do Igbo speakers trust ASR that strips tones?

Fine-Tuning Strategy: Fixing the Blind Spots

What Kind of Dataset Is Needed?

To address the tonal collapse observed in omniASR-CTC-1B, a fine-tuning dataset should have:

1. Tone-accurate transcriptions: Ground truth text must include all diacritics that mark phonemic tone distinctions (à, è, ì, ọ, ụ, etc.)

2. Minimal pairs coverage: Explicitly include utterances with tonal minimal pairs (akwa/akwà/àkwà/ákwá) to force the model to learn acoustic tone distinctions rather than relying on lexical priors

3. Speaker diversity:

   • 20+ speakers across major Igbo dialects (Owerri, Onitsha, Enugu, Nsukka, Afikpo, etc.)
   • Balanced male/female representation
   • Age range 18-65 to capture generational variation

4. Domain coverage:

   • Natural conversational speech (not read text)
   • Code-switching scenarios (English-Igbo mixing)
   • Cultural terms, place names, and idiomatic expressions
   • Varying speech rates and prosodic patterns

5. Audio quality:

   • Clean recordings (SNR >20dB) for initial fine-tuning
   • Noise-augmented variants for robustness testing
   • 16kHz sampling rate (matching omniASR's input format)

How to Assemble/Find Such a Dataset

Option 1: Crowdsourced Collection (Recommended)

Partner with Nigerian universities and Igbo language organizations to recruit native speakers:

1. Recruitment:

   • Advertise through Igbo cultural associations, university linguistics departments
   • Compensation: $15-20/hour (competitive for Nigerian context)
   • Require: Native speakers with literacy in standard Igbo orthography

2. Recording protocol:

   • Mobile app for audio capture (ensures quality control)
   • Prompt design: Mix of scripted (minimal pairs, proverbs) and spontaneous (describe your day, tell a story)
   • Quality assurance: Native speaker reviewers verify tone accuracy

3. Annotation workflow:

   • Speakers self-transcribe with diacritics (who better than native speakers?)
   • Cross-validation: Second annotator reviews 20% of transcriptions
   • Dispute resolution: Linguist adjudicates disagreements

Option 2: Existing Resources + Enhancement

Leverage existing Igbo corpora but add tone annotations:

1. Start with: IgboAPI, JW300 Igbo corpus, NSC Igbo Bible
2. Problem: These lack audio or have incomplete diacritics
3. Solution: Commission native speakers to:
   • Record audio for existing text
   • Add missing tone marks to transcriptions
   • Validate tone-text alignment

Option 3: Hybrid Approach

• Core dataset: 100 hours crowdsourced (Option 1)
• Augmentation: 50 hours from enhanced existing resources (Option 2)
• Validation set: Held-out recordings from this diagnostic dataset

Dataset Size Estimates

These are rough planning estimates by the author, not figures taken from a specific paper:

Minimum viable:

• 10-20 hours of tone-annotated audio
• ~5,000-10,000 utterances (average 5-10 seconds each)
• Cost: ~$5,000-8,000 (recording + transcription + QA)

Recommended for robust performance:

• 50-100 hours of tone-annotated audio
• ~25,000-50,000 utterances
• 20+ speakers (minimum 2 hours each)
• Cost: ~$20,000-35,000

Why this size?

• Our diagnostic shows systematic failure, not just data scarcity, so targeted examples of minimal pairs are more important than raw hours

Bootstrapping strategy:

1. Phase 1 (10 hours): Collect minimal pairs + high-frequency words
2. Evaluate: Measure diacritic accuracy improvement
3. Phase 2 (40 hours): Expand to conversational speech if Phase 1 shows promise
4. Iterate: Continue only if gains justify cost

Expected Outcomes

A 10-20 hour targeted fine-tuning dataset may substantially reduce tonal diacritic loss; empirical validation would be required to confirm specific performance gains. Our monotone clip (File 09) is consistent with a text-prior effect that fine-tuning alone may not fully address, but it is a single clip. Acoustic analysis (F0 tracking) would clarify this before investing in large-scale data collection.

Data Collection Ethics

• Informed consent: Recordings made by the author with full knowledge of public release
• Privacy: All recordings are self-recorded by the author. No third-party identifiable information included.
• Cultural sensitivity: Proverbs and idioms are common knowledge, not sacred/restricted content
• Community benefit: Dataset released open-source to benefit Igbo NLP research
• No exploitation: Zero-compensation labor issue does not apply (self-recorded by community member)

This dataset follows guidelines from the ACM Code of Ethics and Professional Conduct for responsible AI research.

Citation

If you use this dataset, please cite:

@misc{obasi2026igbo,
  title={Igbo Blind Spot Dataset for omniASR-CTC-1B: Systematic Evaluation of Tonal Diacritic Loss},
  author={Obasi, Chizoba},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/Chiz/omniASR-igbo-blindspots}},
  note={Model evaluated: facebook/omniASR-CTC-1B (975M parameters)}
}

References

Choi, A. S. G., & Choi, H. (2025). Fairness of automatic speech recognition: Looking through a philosophical lens. Proceedings of the Eighth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2025), 8(1), 605–614. https://doi.org/10.1609/aies.v8i1.36574

Ezeani, I. (2019). Corpus-based approaches to Igbo diacritic restoration [Doctoral dissertation, The University of Sheffield]. White Rose eTheses Online. https://arxiv.org/abs/2601.18380

Ezeani, I., Hepple, M., Onyenwe, I., & Enemouh, C. (2018). Igbo diacritic restoration using embedding models. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop, 54–60. https://doi.org/10.18653/v1/N18-4008

Harris, C., Mgbahurike, C., Kumar, N., & Yang, D. (2024). Modeling gender and dialect bias in automatic speech recognition. Findings of the Association for Computational Linguistics: EMNLP 2024, 15166–15184. https://doi.org/10.18653/v1/2024.findings-emnlp.890

Koenecke, A., Nam, A., Lake, E., Nudell, J., Quartey, M., Mengesha, Z., Toups, C., Rickford, J. R., Jurafsky, D., & Goel, S. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689. https://doi.org/10.1073/pnas.1915768117

Liang, S., & Levow, G.-A. (2025). Breaking the transcription bottleneck: Fine-tuning ASR models for extremely low-resource fieldwork languages. Proceedings of the Fourth Workshop on NLP Applications to Field Linguistics, 26–37. https://aclanthology.org/2025.fieldmatters-1.3/

Nakatumba-Nabende, J., Kagumire, S., Kantono, C., & Nabende, P. (2025). A systematic literature review on bias evaluation and mitigation in automatic speech recognition models for low-resource African languages. ACM Computing Surveys, 58(4), Article 105. https://doi.org/10.1145/3769089

Nigatu, H. H., Tonja, A. L., Rosman, B., Solorio, T., & Choudhury, M. (2024). The Zeno’s paradox of ‘low-resource’ languages. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 17753–17774. https://doi.org/10.18653/v1/2024.emnlp-main.983

Obasi, C. (2026). Igbo blind spot dataset for omniASR-CTC-1B: Systematic evaluation of tonal diacritic loss [Dataset]. Hugging Face. https://huggingface.co/datasets/Chiz/omniASR-igbo-blindspots

Omnilingual ASR Team, Keren, G., Kozhevnikov, A., Meng, Y., Ropers, C., Setzler, M., Wang, S., Adebara, I., Auli, M., Balioglu, C., Chan, K., Cheng, C., Chuang, J., Droof, C., Duppenthaler, M., Duquenne, P.-A., Erben, A., Gao, C., Mejia Gonzalez, G., Lyu, K., Miglani, S., Pratap, V., Sadagopan, K. R., Saleem, S., Turkatenko, A., Ventayol-Boada, A., Yong, Z.-X., Chung, Y.-A., Maillard, J., Moritz, R., Mourachko, A., Williamson, M., & Yates, S. (2025). Omnilingual ASR: Open-source multilingual speech recognition for 1600+ languages. arXiv. https://arxiv.org/abs/2511.09690

License

• Dataset (audio, transcripts, annotations, metadata, model outputs): CC BY 4.0 (attribution required)
• Code: MIT License (repository)
• Model evaluated: facebook/omniASR-CTC-1B, Apache License 2.0

All recordings are of the dataset author, who consents to their release under CC BY 4.0. The recordings include the author's first name; other names spoken in the recordings are generic example names. No other personal information is included.

Contact

• Author: Chizoba Obasi
• HuggingFace: hf.co/Chiz